dsh-vision-bridge

When your inference model is text-only but your daily workflow keeps throwing images at it, dsh-vision-bridge fills that gap. It registers no extra wrapper provider; instead it reuses the adapter and attachment channel of a DSH provider you already configured, or accepts a custom OpenAI-compatible endpoint you plug in directly. Switching between the two modes is a single dropdown in the settings page, and changes take effect immediately without a restart.

Two pathways do the heavy lifting. On the model side, the vision_bridge_read_image tool returns structured evidence across three layers: OCR text, layout, and semantic description, on demand. On the user side, pasting an image automatically uploads it and inserts a compact one-image-one-ID short reference that works in TEXTAREA, INPUT, and contenteditable fields alike. The settings page adds a one-click connectivity test that sends a sample image and returns latency plus a response snippet, so troubleshooting never requires digging through logs. API keys are resolved through the DSH credentials service, are never echoed back, and the plugin has zero visibility into a DSH provider's base URL or secret.

This plugin is for anyone running a text-only model inside DSH who needs to interpret screenshots, charts, photos, or text embedded in images. It requires DSH 0.1.1+ and Node.js 18+, runs on macOS and Linux, and is released under the MIT license.